Bayesian Neural Networks for Direct Short-Term Net Load Forecasting in Solar-Integrated Residential Buildings

Short-term net load forecasting (STNLF) is of utmost importance for the efficient operation and control of residential buildings, with integrated photovoltaic (PV) systems. The intermittent PV generation increases the uncertainty of STNLF and therefore the error. This study presents a machine learning model using Bayesian neural networks (BNNs) for direct STNLF in solar-integrated residential buildings. Historical data from PV-integrated households in Cyprus, with self-sufficiency between 88.59% and 294.12%, were used for assessing the performance of the BNN model. The proposed model was applied to data from 68 individual and aggregated households in Cyprus. The results indicated that the machine learning model exhibited normalized root mean square error of 7.20% for the aggregated data, rendering it suitable for STNLF in residential buildings. Finally, the model was compared to a naïve persistence model, achieving a skill score of 46.37% at the aggregated household level.